<p>Accurate rice-field mapping is critical for supporting agricultural monitoring in tropical regions where cloud cover and heterogeneous field conditions often limit optical-based observations. However, the capability of different Recurrent Neural Network (RNN) variants to model temporal backscatter dynamics from multi-temporal C-band SAR data, particularly in contrasting topographic settings, remains insufficiently assessed. This study systematically evaluates the performance of Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), and Gated Recurrent Unit (GRU) architectures for rice-field identification in Indramayu Regency, Indonesia, using backscatter and Grey-Level Co-occurrence Matrix (GLCM) texture features statistics derived from VH and VV Sentinel-1 time-series data. Two feature schemes were examined (VH-based and VV-based), and Artificial Neural Network (ANN) as well as one-dimensional Convolutional Neural Network (1D-CNN) were included as benchmark deep learning models under a unified training and validation framework. Results show that the GRU model consistently achieved the highest accuracy across both schemes, reaching an Overall Accuracy (OA) of 0.97 in flat irrigated areas, with superior recall (0.98) and F1-score (0.95) when VH polarization was used, indicating a strong ability to reduce omission errors. GRU also demonstrated the most efficient convergence. Field verification confirmed accurate detection of rice fields across both vegetative and generative growth stages. When applied to more undulating terrain, the GRU model maintained robust performance (OA = 0.94–0.95), despite expected reductions due to terrain-induced SAR distortions. These findings highlight the advantages of GRU-based temporal modeling for operational, weather-independent paddy-field mapping and underscore its potential for scalable rice-monitoring applications in complex agro-ecosystems.</p>